blog

When Agents Prefer Hacking To Failure: Evaluating Misalignment Under Pressure

An AI safety experiment analyzing evaluation hacking in language-model agents, combining controlled task design, behavioral analysis under pressure, and systematic testing across models. Read the write-up for experimental setup, findings, and implications for alignment. (Context: CS 2881r Seminar)

Read more →

Capy Running Coach

An AI-powered running coach on WhatsApp, combining RAG, fine-tuned LLMs, vector search, and a GKE deployment. Read the write-up for system design and results.

Read more →

Helper Brasil Conta Comigo

Oct 29, 2020

Brasil Conta Comigo is a Governmental program that aims to increase the number of professionals in the fight against COVID-19. It intends to do so by recruiting students from the 5º and 6º years of Medical School along with last year students from the Nursing, physiotherapy, and farmaceutical school.

Read more →

Medway Student Allocation

Aug 08, 2020

Medway João Pessoa is an ENEM preparatory program known for its personalized advisory system, where students receive weekly guidance from experienced former ENEM students. This project focuses on improving the student–advisor allocation process by increasing compatibility and balancing workloads, with the goal of reducing advisor burnout and enhancing the experience for both students and advisors.

Read more →